Quote from: Yakystockero on May 15, 2023, 03:07

And personally I have not seen any cases of plagiarism by AIs yet.

...
Now you have.

https://gizmodo.com/ai-art-generators-ai-copyright-stable-diffusion-1850060656

what a surprise!  you take a very specific prompt with a tiny training set you esigned & you generate a similar image (which would not be possible with MJ or DALLE since they dont allow such names to be used). 

 the article in the link is highly misleading - when you read the actual article, you discover they did not generate this image from MJ et al - they were on a very specific, biased snipe hunt to prove their hypothesis (not unreasonable since they were trying to show feasibility) -- they used the original images, did their own training, and then generated images with their own generator.-.  their methodology was specifically looking to generate 'memorized' images and is highly unlikely to occur in the real world - while they show it's theoretically possible for this to occur, they haven't shown it does with existing public generators

we still havent seen such results with publicly available

btw, as previously reported, when i tested a very specific prompt for which i had 80% of the images on SS, the result was competely different from the training images.

This has clearly nothing to do with AI images creations, and
It simply seems a low quality jpg, or overcompressed image.

Did you even read the article? How does this have "clearly nothing to do with AI image creations", when the image on the right WAS created with an image AI creator, Stable Diffusion, to be exact?  It's not a low quality compressed image, it is the image Stable Diffusion AI created.

And Stable Diffusion has not been given "one image of a women", but all of the internet. It's currently one of the leading AI image creators out there.
Yes, in this case the people achieving this result gave it a prompt that equaled the exact dataset the AI was trained with, this was not a chance result. No one claimed they entered "blond woman" into the prompt field and by chance got a result that looked exactly like an existing photo. But the point of this is to show the problem that AI HAS the capacity to completely recreate an image that belongs to someone else.

 
Here's the actual paper:

https://arxiv.org/pdf/2301.13188.pdf

When reading bear in mind all the major AI generators now use diffusion models rather than GANs. I described the method before and was told "that's not how AI works" by someone who then proceeded to describe GANs.

The dataset they used to extract images is Stable Diffusion "the largest and most popular open-source diffusion model. This model is an 890 million parameter text-conditioned diffusion model trained on 160 million images."

Here's an extract:
"Examples of the images that we extract from Stable Diffusion v1.4 using random sampling and our membership inference procedure. The top row shows the original images and the bottom row shows our extracted images.
"

Here is what they actually use the smaller data set to investigate (in the second part of the paper):

"The above experiments are visually striking and clearly indicate that memorization is pervasive in large diffusion models�and that data extraction is feasible. But these experiments do not explain why and how these models memorize training data."

So they didn't use a hand picked dataset to prove they could extract the images at all. They used the largest set available. Only subsequently did they use a smaller set to show how they were stored.

 
Here's the actual paper:

https://arxiv.org/pdf/2301.13188.pdf

When reading bear in mind all the major AI generators now use diffusion models rather than GANs. I described the method before and was told "that's not how AI works" by someone who then proceeded to describe GANs.

The dataset they used to extract images is Stable Diffusion "the largest and most popular open-source diffusion model. This model is an 890 million parameter text-conditioned diffusion model trained on 160 million images."

Here's an extract:
"Examples of the images that we extract from Stable Diffusion v1.4 using random sampling and our membership inference procedure. The top row shows the original images and the bottom row shows our extracted images.
"

Here is what they actually use the smaller data set to investigate (in the second part of the paper):

"The above experiments are visually striking and clearly indicate that memorization is pervasive in large diffusion models�and that data extraction is feasible. But these experiments do not explain why and how these models memorize training data."

So they didn't use a hand picked dataset to prove they could extract the images at all. They used the largest set available. Only subsequently did they use a smaller set to show how they were stored.

I read the paper: while aiming specifically at Stable Diffusion (with images based on the LAION dataset) using a specific algorithm  they managed to "extract" 50 images out of 175 million, and all those 50 images were duplicated at least 100 times in the dataset. In order to retrieve the images they had to use as a prompt a string siphoned from the LAION dataset itself. Speaking about doctored stats...

One would conclude that, unless you're specifically hunting for a scandal, the probability of getting a "tainted" image from Stable Diffusion is 1 over 3,500,000.
An average human life lasts (with a bit of luck) about 29,000 days, hence - statistically speaking - you'd have to create with Stable Diffusion 1206 images a day (starting on the day of your birth) before getting one.  Roll up your sleeves... 

The point was to demonstrate that the images are stored in some form which was demonstrated by retrieving them with the right prompt. The relevant point is that the images are, in fact, compressed in a database that is pulled from. I am on mobile so can't  get to the paper right now. Doesn't it conclude that something like 0.03% of the time they were getting recognisable results, so almost 1 in 300? (tough this is really irrelevant, the point is they conclusively demonstrated that original images are stored, how good the app is at covering it up is less relevant).

ETA. Just reread your post. Lol made up stats indeed. You took the total database size and divided it by the 50 images to calculate the probability? Come on, you're a scientist.

Oh, about the stats please teach me! And about the storage, please consider recommending Stable Diffusion not to keep their disks in a mouldy cellar and use the cloud instead